Neural Architecture Search (NAS) for Human Pose Estimation.
Search part-specific Cell-based Neural Fabrics (CNFs) with the guide of prior knowledge of human body structure.
Introduction
Neural Architecture Search (NAS), the process of learning the structure of neural network, can play a potential role at automatically designing network architectures. Current methods mainly take image classification as a basic task and only search for a micro cell to build a chain-like structure, thus the neural search space is still at the limit of a micro search space. However, when applying NAS to dense prediction tasks such as semantic segmentation and human pose estimation, the micro search space is no longer able to generate more complex architectures. Therefore, it become a necessity to artificially design the macro search space allowing identifying hierarchical structure upon cells for these tasks. In addition, existing works focus on discovering an alternative to the human-designed module in a common pipeline. Such practice actually decouples the automating architecture engineering from tasks, and is thus unable to take advantage of the domain knowledge of a specific task.
In this work, we study how to search neural architectures with the guide of prior knowledge for human pose estimation task and propose a framework named Pose Neural Fabrics Search. We notice that modern methods conducting human pose estimation based on deep CNNs, regardless of top-down or bottom-up pipeline, convert it into pixel-wise prediction problem; they usually focus on two aspects: neural architecture design and pose representation. Next, we will discuss our motivations from these two aspects. ...